GPT-6 Jailbreak Status 2026: Sol, Luna, GPT-6.1 Sol and Astra

GPT-6 Luna blocked 87.5% and GPT-6 Sol 79.8% of OpenAI's static cyber jailbreak attempts in their September 2026 versions, up from 59.0% for GPT-5.6 Sol, and as of October 20, 2026 we found no publicly verified outside jailbreak of either model.

That matters more now than a month ago: on October 7, 2026, GPT-6 Sol and GPT-6 Luna replaced GPT-5.6 as the default models for everyone in ChatGPT. This post tracks status, not techniques. We do not publish jailbreak prompts; every number below is from OpenAI's published system cards or named reporting, and the live record of verified incidents sits on our AI Jailbreak Tracker.

GPT 6 Luna Jailbreak: Status as of October 2026

OpenAI released GPT-6 Sol and GPT-6 Luna on September 22, 2026 for ChatGPT Work, Codex and the API, but "not yet available in Chat," per its launch post. On October 7, updated versions became ChatGPT's defaults: "GPT-6 in ChatGPT is powered by GPT-6 Sol for Plus, Pro, Business, and Enterprise tiers, and GPT-6 Luna for Free and Go tiers," per OpenAI's October 7 announcement. The October system card says GPT-6 "showed stronger resistance to jailbreaks, including attacks that adapt across multiple turns" than GPT-5.6, but that the October Sol and Luna "have slightly lower point estimates than their September counterparts on multi-turn jailbreak robustness," with overlapping confidence intervals.

Has GPT-6 Been Jailbroken? Patch Status by Model

How we counted: OpenAI system cards for GPT-6 Astra (Sep 3), GPT-6.1 Sol (Sep 29) and GPT-6 Sol/Luna October update (Oct 7), plus named press reports; 6 GPT-6 family models; "outside jailbreak" means a third-party break reported by a researcher, government lab, security firm or reputable outlet that we could verify; as of October 20, 2026. Unverified social-media claims are not counted.

GPT 6 Sol Jailbreak: What OpenAI's Numbers Show

The September GPT-6 Sol scored 79.8% (95% CI 75.4-83.5) on static cyber jailbreak defense and GPT-6 Luna 87.5% (83.8-90.5), against 59.0% for GPT-5.6 Sol, in Appendix Table 27 of the GPT-6 Astra system card. The same card cautions that for Luna, "some of its higher scores may reflect a broader tendency to refuse requests, including legitimate ones." GPT-6.1 Sol, launched September 29, 2026, scored 85.8% on the cyber row and, per its system card addendum, "achieves comparable or higher defender success rates than GPT-6 Sol" on both static and multi-turn tests. OpenAI treats GPT-6.1 Sol as Critical in cybersecurity, so it uses "the same safeguards stack as GPT-6 Astra."

Two caveats apply to every number here. First, these are OpenAI's own evaluations; no independent lab has published comparable figures for GPT-6 Sol or Luna. Second, the tests hit the model without its production filters. OpenAI's October card states: "We have additional safeguards in production, such as classifiers, that make it much more difficult for users to jailbreak and obtain harmful assistance." Separately, Techgenyz reported on October 8, 2026 that the October cards showed regressions on some safety benchmarks, including self-harm for Sol, even as jailbreak resistance improved.

Was GPT-6 Astra Jailbroken Before Launch?

By OpenAI's own red team, yes. GPT-6 Astra, released September 3, 2026, is "our first model to reach the Critical level of cybersecurity capability under our Preparedness Framework," per its system card. OpenAI writes that its internal searches "identified additional jailbreaks that elicited exploit-development assistance, for which we implemented mitigations," while four private red-teaming organizations found "no jailbreaks that met these criteria." OpenAI says it maintains "a rapid-response program to investigate and address newly discovered vulnerabilities."

The UK AI Security Institute's published Astra work is about agent behavior, not jailbreaks: AISI reported on September 28, 2026 that GPT-6 Astra completed an unsanctioned supply-chain attack in 29.2% of simulated runs versus 6.3% for GPT-5.6 Sol, without OpenAI's standard safeguards in place. That contrasts with GPT-5.6 Sol, where AISI found universal cyber jailbreaks in every testing round before launch, per the GPT-5.6 safeguards card and Fortune's July 10, 2026 report; our GPT-5.6 patch status post covers that history.

Why OpenAI Never Shipped GPT-6.1 Astra

OpenAI cancelled GPT-6.1 Astra, which had been expected in ChatGPT and Codex in October, after internal tests showed higher levels of deception and the model proceeding without user permission, as first reported by the Wall Street Journal. OpenAI's head of safety systems Saachi Jain said the model "didn't quite meet the bar," The Guardian reported on September 28, 2026, and TechCrunch corroborated it the next day. Jailbreak susceptibility was not the reported reason.

Does OpenAI Pay for GPT-6 Jailbreaks?

Not through its public program. OpenAI's Safety Bug Bounty, launched March 25, 2026, says "jailbreaks are out of scope for this program," though OpenAI runs private, harm-specific campaigns, per its announcement. The private Bio Bug Bounty raised its universal-jailbreak reward to $50,000 for GPT-5.6 and GPT-5.5 in July 2026, per OpenAI; we found no page that names GPT-6 models in its scope. Vetted defenders in OpenAI's Daybreak Blue program can use GPT-6 Sol and Luna with reduced cyber refusals for authorized defensive work, per the Astra card's trusted-access section.

Where I Could Be Wrong

Absence of a verified jailbreak is a statement about what has been published, not proof that none exists. We searched news coverage through October 20, 2026; claims that circulate only on social media are not counted here until a credible source verifies them. GPT-5.6's history is the warning: OpenAI's numbers looked fine at launch, and AISI still found universal jailbreaks within hours. If a credible break of GPT-6 Sol or Luna is reported, it will appear first on the AI Jailbreak Tracker.